Skip to content

CUDA: loop over ne2*ne3 in case it overflows #19538

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
am17an merged 2 commits into from
Feb 13, 2026
Merged

CUDA: loop over ne2*ne3 in case it overflows #19538

Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
3b93d39
CUDA: loop over ne2*ne3 in case it overflows
am17an Feb 10, 2026
0fd79e5
use fastdiv
am17an Feb 12, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
56 changes: 32 additions & 24 deletions ggml/src/ggml-cuda/convert.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,31 +8,34 @@
template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __restrict__ y,
const int64_t ne00, const int64_t ne01, const int64_t ne02,
const int64_t s01, const int64_t s02, const int64_t s03) {
const int64_t ne0203, const int64_t s01, const int64_t s02, const int64_t s03) {
const int64_t i00 = 2 * (int64_t(blockDim.x)*blockIdx.x + threadIdx.x);

if (i00 >= ne00) {
return;
}

const int64_t i01 = blockIdx.y;
const int64_t i02 = blockIdx.z % ne02;
const int64_t i03 = blockIdx.z / ne02;

const int64_t ibx0 = i03*s03 + i02*s02 + i01*s01;
for (int64_t i0203 = blockIdx.z; i0203 < ne0203; i0203 += gridDim.z) {
const int64_t i02 = i0203 % ne02;
const int64_t i03 = i0203 / ne02;

@JohannesGaessler JohannesGaessler Feb 12, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Preferably use fastdiv here.


const int64_t ib = ibx0 + i00/qk; // block index
const int64_t iqs = (i00%qk)/qr; // quant index
const int64_t iybs = i00 - i00%qk; // y block start index
const int64_t y_offset = qr == 1 ? 1 : qk/2;
const int64_t ibx0 = i03*s03 + i02*s02 + i01*s01;

// dequantize
float2 v;
dequantize_kernel(vx, ib, iqs, v);
const int64_t ib = ibx0 + i00/qk; // block index
const int64_t iqs = (i00%qk)/qr; // quant index
const int64_t iybs = i00 - i00%qk; // y block start index
const int64_t y_offset = qr == 1 ? 1 : qk/2;

const int64_t iy0 = ((i03*ne02 + i02)*ne01 + i01)*ne00 + iybs + iqs;
y[iy0 + 0] = ggml_cuda_cast<dst_t>(v.x);
y[iy0 + y_offset] = ggml_cuda_cast<dst_t>(v.y);
// dequantize
float2 v;
dequantize_kernel(vx, ib, iqs, v);

const int64_t iy0 = ((i03*ne02 + i02)*ne01 + i01)*ne00 + iybs + iqs;
y[iy0 + 0] = ggml_cuda_cast<dst_t>(v.x);
y[iy0 + y_offset] = ggml_cuda_cast<dst_t>(v.y);
}
}

template <bool need_check>
Expand Down Expand Up @@ -485,9 +488,10 @@ template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
static void dequantize_block_cuda(const void * vx, dst_t * y,
const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
const int64_t s01, const int64_t s02, const int64_t s03, cudaStream_t stream) {
const dim3 num_blocks((ne00 + 2*CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / (2*CUDA_DEQUANTIZE_BLOCK_SIZE), ne01, ne02*ne03);
const int64_t ne0203 = ne02*ne03;
const dim3 num_blocks((ne00 + 2*CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / (2*CUDA_DEQUANTIZE_BLOCK_SIZE), ne01, (int)std::min(ne0203, (int64_t)65535));
dequantize_block<qk, qr, dequantize_kernel><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>
(vx, y, ne00, ne01, ne02, s01, s02, s03);
(vx, y, ne00, ne01, ne02, ne0203, s01, s02, s03);
}

template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
Expand Down Expand Up @@ -613,31 +617,35 @@ static void dequantize_row_mxfp4_cuda(const void * vx, dst_t * y, const int64_t
template <typename src_t, typename dst_t>
static __global__ void convert_unary(
const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t ne00, const int64_t ne01, const int64_t ne02,
const int64_t s01, const int64_t s02, const int64_t s03) {
const int64_t ne0203, const int64_t s01, const int64_t s02, const int64_t s03) {
const int64_t i00 = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;

if (i00 >= ne00) {
return;
}

const int64_t i01 = blockIdx.y;
const int64_t i02 = blockIdx.z % ne02;
const int64_t i03 = blockIdx.z / ne02;

const src_t * x = (const src_t *) vx;

const int64_t ix = i03*s03 + i02*s02 + i01*s01 + i00;
const int64_t iy = ((i03*ne02 + i02)*ne01 + i01)*ne00 + i00;
y[iy] = ggml_cuda_cast<dst_t>(x[ix]);
for (int64_t i0203 = blockIdx.z; i0203 < ne0203; i0203 += gridDim.z) {
const int64_t i02 = i0203 % ne02;
const int64_t i03 = i0203 / ne02;

@JohannesGaessler JohannesGaessler Feb 12, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Same as above.


const int64_t ix = i03*s03 + i02*s02 + i01*s01 + i00;
const int64_t iy = ((i03*ne02 + i02)*ne01 + i01)*ne00 + i00;
y[iy] = ggml_cuda_cast<dst_t>(x[ix]);
}
}

template <typename src_t, typename dst_t>
static void convert_unary_cuda(const void * vx, dst_t * y,
const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
const int64_t s01, const int64_t s02, const int64_t s03, cudaStream_t stream) {
const dim3 num_blocks((ne00 + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE, ne01, ne02*ne03);
const int64_t ne0203 = ne02*ne03;
const dim3 num_blocks((ne00 + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE, ne01, (int)std::min(ne0203, (int64_t)65535));
convert_unary<src_t><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>
(vx, y, ne00, ne01, ne02, s01, s02, s03);
(vx, y, ne00, ne01, ne02, ne0203, s01, s02, s03);
}

template <typename src_t, typename dst_t>
Expand Down
Loading